@contentgrowth/llm-service 0.6.99 → 0.7.1

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@contentgrowth/llm-service",
-  "version": "0.6.99",
+  "version": "0.7.1",
   "description": "Unified LLM Service for Content Growth",
   "main": "src/index.js",
   "type": "module",

package/src/llm/providers/gemini-provider.js

@@ -90,6 +90,26 @@ export class GeminiProvider extends BaseLLMProvider {
                 case 'user':
                     role = 'user';
                     parts = [{ text: msg.content }];
+
+                    // Enhancement: If this is the LAST message (current turn), append the reminder.
+                    // This helps the model respect the system prompt (especially format) even with long context history.
+                    if (index === geminiMessages.length - 1) {
+                        let reminder = "";
+                        if (options.responseFormat === 'json' || options.responseFormat?.type === 'json_schema' || options.responseSchema) {
+                            reminder = "\n\n[SYSTEM NOTE: The output MUST be valid JSON as per the schema. Do not include markdown formatting or explanations.]";
+                        } else {
+                            reminder = "\n\n[SYSTEM NOTE: Please ensure your response adheres strictly to the constraints defined in the System Prompt.]";
+                        }
+
+                        // Append to the existing text part (Safest method)
+                        const lastPart = parts.find(p => p.text);
+                        if (lastPart) {
+                            lastPart.text += reminder;
+                        } else {
+                            // Fallback if message was image-only
+                            parts.push({ text: reminder });
+                        }
+                    }
                     break;
                 case 'assistant':
                     role = 'model';
@@ -111,6 +131,16 @@ export class GeminiProvider extends BaseLLMProvider {
                             response: { content: msg.content },
                         }
                     }];
+
+                    // Fix for JSON mode: If JSON is requested, remind the model to output JSON after tool execution
+                    // This is necessary because strict JSON mode is disabled when tools are present.
+                    if (options.responseFormat === 'json' || options.responseFormat?.type === 'json_schema' || options.responseSchema) {
+                        parts.push({ text: "\n\n[SYSTEM NOTE: The output MUST be valid JSON as per the schema. Do not include markdown formatting or explanations.]" });
+                    } else {
+                        // Generic reminder to help model stay on track with system prompt instructions (e.g. formatting)
+                        // even if no specific JSON mode is configured.
+                        parts.push({ text: "\n\n[SYSTEM NOTE: Please ensure your response adheres strictly to the constraints defined in the System Prompt.]" });
+                    }
                     break;
                 default:
                     return null;

package/src/llm-service.js

@@ -159,18 +159,19 @@ export class LLMService {
         }
     }
 
-    async chatWithTools(messages, tenantId, systemPrompt, tools = []) {
+    async chatWithTools(messages, tenantId, systemPrompt, tools = [], options = {}) {
         const provider = await this._getProvider(tenantId);
 
         let currentMessages = [...messages];
         const MAX_ITERATIONS = 10; // Prevent infinite loops
         let iteration = 0;
 
-        // Initial call - no options since tools + JSON mode are incompatible
+        // Initial call
         const initialResponse = await provider.chatCompletion(
             currentMessages,
             systemPrompt,
-            tools
+            tools,
+            options
         );
 
         let { content, tool_calls, parsedContent } = initialResponse;
@@ -184,11 +185,12 @@ export class LLMService {
             // Execute tools using the provider's helper (which formats results for that provider)
             await provider.executeTools(tool_calls, currentMessages, tenantId, this.toolImplementations, this.env);
 
-            // Next call - no options
+            // Next call
             const nextResponse = await provider.chatCompletion(
                 currentMessages,
                 systemPrompt,
-                tools
+                tools,
+                options
             );
 
             content = nextResponse.content;